Low-level and high-level prior learning for visual saliency estimation

Abstract

Visual saliency estimation is an important issue in multimedia modeling and computer vision, and constitutes a research field that has been studied for decades. Many approaches have been proposed to solve this problem. In this study, we consider the visual attention problem blue with respect to two aspects: low-level prior learning and high-level prior learning. On the one hand, inspired by the concept of chance of happening, the low-level priors, i.e., Color Statistics-based Priors (CSP) and Spatial Correlation-based Priors (SCP), are learned to describe the color distribution and contrast distribution in natural images. On the other hand, the high-level priors, i.e., the relative relationships between objects, are learned to describe the conditional priority between different objects in the images. In particular, we first learn the low-level priors that are statistically based on a large set of natural images. Then, the high-level priors are learned to construct a conditional probability matrix blue that reflects the relative relationship between different objects. Subsequently, a saliency model is presented by integrating the low-level priors, the high-level priors and the Center Bias Prior (CBP), in which the weights that correspond to the low-level priors and the high-level priors are learned based on the eye tracking data set. The experimental results demonstrate that our approach outperforms the existing techniques.